Claude Cowork and Claude Code Have Quietly Become My Universal Front End. More Than Excel, Jupyter or Colab.

Slide 1 - I'm Claude Cowork

Hand me your data and a question, and I'll pull it, clean it, analyze it, build a model, and write the whole thing up as a report or PDF. There's a good deal more I can do, too. Let me show you...

Slide 2 - Chat could only talk. I can act.

A chat assistant answers questions inside a box. As your coworker, I reach out into the real world: your files, live data, the web, your apps, and finish the job, not just describe it.

Read your files · pull live data · run analysis · write reports.

How: Talk to me the way you already do, and I do the work.

Slide 3 - I sit on your desktop, beside your chat.

I'm a desktop app, not just a browser tab. Two ways to work side by side: Cowork for everyday tasks, and Claude Code when something needs to run on your machine.

How: Get the Claude desktop app, open Cowork, and start a session.

Slide 4 - Just posted to the FRED API? Already on my desk.

Point me at an API, and the moment a number lands, I pull it and chart it. US retail sales just posted to the FRED API: a record $763.7B, up 0.9% on the month. With inflation at 4.2%, plenty of that is higher prices, not extra spending.

How: Point me to the API docs. Public feeds like FRED or Yahoo Finance often need no key; for others, paste the key and allow the domain in Settings.

Slide 5 - I reach your data backends, too.

I can reach your own systems two ways: straight through your API, or via an MCP server. Point me at tigzig.com's Tremor and I pull live data from the back end.

How: Give me the API docs, or add your MCP server as a connector in Settings. Then just ask.

Slide 6 - I work your files where they live.

Give me access to a folder and I read, edit, rename and organise what's in it, right on your drive. The work stays where you keep it, ready to use.

Read & edit · rename & sort · tidy folders · save final output.

How: Pick the folders I'm allowed to touch. I'll work inside them.

Slide 7 - Hand me a messy file. Get back a clean report.

A buggy spreadsheet landed in your inbox and you need an answer fast. I clean it, check it, find the story, and hand you a tidy report. You review the numbers, we refine, the analyst still owns the call.

Fix broken rows · spot bad data · de-duplicate · report out.

How: Drop the file in a folder I can see, and tell me what you want from it.

Slide 8 - Big data doesn't go to my head.

An 11.85-million-row, 1.6GB file? Read and analyzed in 13 seconds, with so much headroom left I could take more than double that. Only the results come back to me, never the whole file, so the file's size is not the limit.

How: Just point me at the file. Past a few hundred MB I switch to querying it directly, instead of loading the whole thing into memory.

Slide 9 - I don't just fetch. I analyze.

I explore before I conclude, cutting the data every way to see what's really going on, and I show you the results as I go. Here I overlaid two views of US unemployment to surface what one chart alone would hide.

How: Tell me the question. I'll dig in and show you results as I go.

Slide 10 - I don't stop at charts. I build models.

Point me at a target and I train on the whole file, then validate it on held-out data.

Gradient boosting · scikit-learn · ROC, gains, confusion.

How: Tell me what to predict. I'll train it, validate it, and show the curves.

Slide 11 - I run the stats, too, and tell you what holds up.

Response rates, conversion rates, A/B tests, the bread and butter of marketing. I run the test and tell you whether the difference is real, not just noise.

How: Give me the two groups. I'll run the test and call it.

Slide 12 - I show my work, and we iterate.

No black box, and no one-shot. I share the exploration first, the missing values, the oddities, the data issues, as a page you can open right in your browser. We fix and refine, then build the model, then tune it together.

EDA & missing values · open it in your browser · we refine together.

How: Ask me to share the exploration as a page you can open in your browser.

Slide 13 - I package it the way you need to send it.

A clean one-pager or a multi-page report, then out as whatever the moment needs, including a PDF with clickable links.

How: Just say the format: HTML, PDF, Word, PowerPoint or Excel.

Slide 14 - I'll build you a live dashboard.

Not just a static report, a page you reopen tomorrow and it pulls fresh data each time. A tracker, a metrics page, a queue you keep coming back to.

Live tracker · metrics page · refreshes itself.

How: Ask for "a live page I can reopen". It pulls fresh data each time.

Slide 15 - I read the real web page, not a snippet.

Give me a link and I open the actual web page and fetch the whole thing with my own tools, the real text, not just a short search summary. So I can quote it properly and cut down on guesswork.

Open the real URL · read the whole page · less guesswork.

How: Give me the link or the topic. I'll open the page and read it.

Slide 16 - I run on a schedule.

Set it once and walk away. A morning briefing, a weekly data refresh, a Monday status check, I'll run it on time and have the result waiting for you.

Daily briefing · weekly refresh · month-end report.

How: Say "every morning at 7" or "every Monday", and I'll set it up.

Slide 17 - I'll drive your browser and desktop.

When a tool has no tidy API, I'll use it the way you would, clicking through a website or a desktop app to get the thing done.

How: Ask me to open a site or app. I'll connect to your browser when you're ready.

Slide 18 - I remember you.

Tell me your style once, your fonts, your colours, the way you like reports, and I carry it forward. Point me at a preferences file in your project, or just say "remember this".

Your style · a preferences file · across sessions.

How: Point me to your preferences file, or say "remember this", and I apply it next time.

Slide 19 - What's in my toolkit.

A private Python sandbox - roughly 4GB of memory and 10GB of disk to work in. Install almost anything from PyPI - the whole Python data and ML stack is open to me.

• Data: pandas, polars and pyarrow, to wrangle files far bigger than memory.
• Models & charts: scikit-learn for machine learning, matplotlib for the visuals.
• Images and more: OpenCV and Pillow for image work, and since almost any package installs, audio and other toolkits are a step away.

The fine print: the very biggest packages (XGBoost, TensorFlow, up to ~600MB) can run past the current 45-second install window and not finish; most install in seconds.

Beyond the sandbox, I plug into your tools - Gmail, Drive, Slack, Notion, and a growing list of data providers, whatever's been made available. And for your own systems, custom connectors through MCP servers.

Slide 20 - What I can't do (and the handoff).

I can't run code directly on your laptop, and I won't move money or place trades. For code that has to run on your machine, my teammate Claude Code handles it: I write the instructions, it executes them.

No code on your machine · no trades or transfers · Claude Code, same team.

How: Claude Code sits right there, in the desktop app or in VS Code. Hand it the note and it runs on your machine.

Slide 21 - How Amar works with me.

Amar runs an AI-agent-first shop at tigzig.com. He points me at data, on Tremor and public APIs, I draft, analyse and build, and Claude Code executes on his machine. I've built him HTML one-pagers, PDF reports and models.

How: Visit tigzig.com for free data and APIs. Want me to explore it? I can reach the site directly from Cowork.

Slide 22 - Resources

Everything above runs in Claude's Cowork mode. The guides, and a couple of good reads:

• Cowork, the product: claude.com/product/cowork
• Getting started with Cowork (best practices): claude.com/blog/best-practices-for-getting-started-with-claude-cowork
• Why HTML beats Markdown, by Thariq Shihipar (Anthropic): claude.com/blog/using-claude-code-the-unreasonable-effectiveness-of-html
• Let Claude use your computer: support.claude.com

tigzig.com - the analyst's tool shed · AI-agent first · 40+ live apps · 200+ build guides · security checklist · MCP & API servers. Point Claude Code or any AI agent at it, and it picks up the work for you.

Slide 23 - By the way, I made this too

This whole deck was built in Cowork as HTML, then exported to a PDF. Same tools, same desk.

Claude Cowork, with Amar Harolikar · tigzig.com